[PATCH 2/2] UBI: Block: Add blk-mq support

From: Richard Weinberger
Date: Sat Jan 10 2015 - 09:30:25 EST


Convert the driver to blk-mq.
Beside of moving to the modern block interface this change boosts
also the performance of the driver.

nand: device found, Manufacturer ID: 0x2c, Chip ID: 0xda
nand: Micron NAND 256MiB 3,3V 8-bit
nand: 256MiB, SLC, page size: 2048, OOB size: 64

root@debian-armhf:~# dd if=/dev/ubiblock0_0 of=/dev/zero bs=1M
243+1 records in
243+1 records out
255080448 bytes (255 MB) copied, 4.39295 s, 58.1 MB/s

vs.

root@debian-armhf:~# dd if=/dev/ubiblock0_0 of=/dev/zero bs=1M
243+1 records in
243+1 records out
255080448 bytes (255 MB) copied, 2.87676 s, 88.7 MB/s

Cc: hch@xxxxxxxxxxxxx
Cc: axboe@xxxxxx
Cc: tom.leiming@xxxxxxxxx
Signed-off-by: Richard Weinberger <richard@xxxxxx>
Tested-by: Ezequiel Garcia <ezequiel@xxxxxxxxxxxxxxxxxxxx>
---
drivers/mtd/ubi/block.c | 213 +++++++++++++++++++++++-------------------------
1 file changed, 104 insertions(+), 109 deletions(-)

diff --git a/drivers/mtd/ubi/block.c b/drivers/mtd/ubi/block.c
index 6b6bce2..bc088ab 100644
--- a/drivers/mtd/ubi/block.c
+++ b/drivers/mtd/ubi/block.c
@@ -42,11 +42,12 @@
#include <linux/list.h>
#include <linux/mutex.h>
#include <linux/slab.h>
-#include <linux/vmalloc.h>
#include <linux/mtd/ubi.h>
#include <linux/workqueue.h>
#include <linux/blkdev.h>
+#include <linux/blk-mq.h>
#include <linux/hdreg.h>
+#include <linux/scatterlist.h>
#include <asm/div64.h>

#include "ubi-media.h"
@@ -67,6 +68,13 @@ struct ubiblock_param {
char name[UBIBLOCK_PARAM_LEN+1];
};

+struct ubiblock_pdu {
+ struct request *req;
+ struct ubiblock *dev;
+ struct work_struct work;
+ struct ubi_sgl usgl;
+};
+
/* Numbers of elements set in the @ubiblock_param array */
static int ubiblock_devs __initdata;

@@ -84,11 +92,10 @@ struct ubiblock {
struct request_queue *rq;

struct workqueue_struct *wq;
- struct work_struct work;

struct mutex dev_mutex;
- spinlock_t queue_lock;
struct list_head list;
+ struct blk_mq_tag_set tag_set;
};

/* Linked list of all ubiblock instances */
@@ -181,31 +188,17 @@ static struct ubiblock *find_dev_nolock(int ubi_num, int vol_id)
return NULL;
}

-static int ubiblock_read_to_buf(struct ubiblock *dev, char *buffer,
- int leb, int offset, int len)
-{
- int ret;
-
- ret = ubi_read(dev->desc, leb, buffer, offset, len);
- if (ret) {
- dev_err(disk_to_dev(dev->gd), "%d while reading from LEB %d (offset %d, length %d)",
- ret, leb, offset, len);
- return ret;
- }
- return 0;
-}
-
-static int ubiblock_read(struct ubiblock *dev, char *buffer,
- sector_t sec, int len)
+static int ubiblock_read(struct ubiblock_pdu *pdu)
{
- int ret, leb, offset;
- int bytes_left = len;
- int to_read = len;
- u64 pos = sec << 9;
+ int ret, leb, offset, bytes_left;
+ int to_read = blk_rq_bytes(pdu->req);
+ struct ubiblock *dev = pdu->dev;
+ u64 pos = blk_rq_pos(pdu->req) << 9;

/* Get LEB:offset address to read from */
offset = do_div(pos, dev->leb_size);
leb = pos;
+ bytes_left = to_read;

while (bytes_left) {
/*
@@ -215,11 +208,10 @@ static int ubiblock_read(struct ubiblock *dev, char *buffer,
if (offset + to_read > dev->leb_size)
to_read = dev->leb_size - offset;

- ret = ubiblock_read_to_buf(dev, buffer, leb, offset, to_read);
- if (ret)
+ ret = ubi_read_sg(dev->desc, leb, &pdu->usgl, offset, to_read);
+ if (ret < 0)
return ret;

- buffer += to_read;
bytes_left -= to_read;
to_read = bytes_left;
leb += 1;
@@ -228,79 +220,6 @@ static int ubiblock_read(struct ubiblock *dev, char *buffer,
return 0;
}

-static int do_ubiblock_request(struct ubiblock *dev, struct request *req)
-{
- int len, ret;
- sector_t sec;
-
- if (req->cmd_type != REQ_TYPE_FS)
- return -EIO;
-
- if (blk_rq_pos(req) + blk_rq_cur_sectors(req) >
- get_capacity(req->rq_disk))
- return -EIO;
-
- if (rq_data_dir(req) != READ)
- return -ENOSYS; /* Write not implemented */
-
- sec = blk_rq_pos(req);
- len = blk_rq_cur_bytes(req);
-
- /*
- * Let's prevent the device from being removed while we're doing I/O
- * work. Notice that this means we serialize all the I/O operations,
- * but it's probably of no impact given the NAND core serializes
- * flash access anyway.
- */
- mutex_lock(&dev->dev_mutex);
- ret = ubiblock_read(dev, bio_data(req->bio), sec, len);
- mutex_unlock(&dev->dev_mutex);
-
- return ret;
-}
-
-static void ubiblock_do_work(struct work_struct *work)
-{
- struct ubiblock *dev =
- container_of(work, struct ubiblock, work);
- struct request_queue *rq = dev->rq;
- struct request *req;
- int res;
-
- spin_lock_irq(rq->queue_lock);
-
- req = blk_fetch_request(rq);
- while (req) {
-
- spin_unlock_irq(rq->queue_lock);
- res = do_ubiblock_request(dev, req);
- spin_lock_irq(rq->queue_lock);
-
- /*
- * If we're done with this request,
- * we need to fetch a new one
- */
- if (!__blk_end_request_cur(req, res))
- req = blk_fetch_request(rq);
- }
-
- spin_unlock_irq(rq->queue_lock);
-}
-
-static void ubiblock_request(struct request_queue *rq)
-{
- struct ubiblock *dev;
- struct request *req;
-
- dev = rq->queuedata;
-
- if (!dev)
- while ((req = blk_fetch_request(rq)) != NULL)
- __blk_end_request_all(req, -ENODEV);
- else
- queue_work(dev->wq, &dev->work);
-}
-
static int ubiblock_open(struct block_device *bdev, fmode_t mode)
{
struct ubiblock *dev = bdev->bd_disk->private_data;
@@ -374,6 +293,67 @@ static const struct block_device_operations ubiblock_ops = {
.getgeo = ubiblock_getgeo,
};

+static void ubiblock_do_work(struct work_struct *work)
+{
+ int ret;
+ struct ubiblock_pdu *pdu = container_of(work, struct ubiblock_pdu, work);
+
+ ret = ubiblock_read(pdu);
+ blk_mq_end_request(pdu->req, ret ?: 0);
+}
+
+static int ubiblock_queue_rq(struct blk_mq_hw_ctx *hctx,
+ const struct blk_mq_queue_data *bd)
+{
+ int ret;
+ struct request *req = bd->rq;
+ struct ubiblock *dev = hctx->queue->queuedata;
+ struct ubiblock_pdu *pdu = blk_mq_rq_to_pdu(req);
+
+ if (req->cmd_type != REQ_TYPE_FS)
+ return BLK_MQ_RQ_QUEUE_ERROR;
+
+ if (blk_rq_pos(req) + blk_rq_cur_sectors(req) >
+ get_capacity(req->rq_disk))
+ return BLK_MQ_RQ_QUEUE_ERROR;
+
+ if (rq_data_dir(req) != READ)
+ return BLK_MQ_RQ_QUEUE_ERROR; /* Write not implemented */
+
+ pdu->usgl.list_pos = 0;
+ pdu->usgl.page_pos = 0;
+
+ blk_mq_start_request(req);
+ ret = blk_rq_map_sg(hctx->queue, req, pdu->usgl.sg);
+
+ queue_work(dev->wq, &pdu->work);
+
+ return BLK_MQ_RQ_QUEUE_OK;
+}
+
+static int ubiblock_init_request(void *data, struct request *req,
+ unsigned int hctx_idx,
+ unsigned int request_idx,
+ unsigned int numa_node)
+{
+ struct ubiblock *dev = data;
+ struct ubiblock_pdu *pdu = blk_mq_rq_to_pdu(req);
+
+ pdu->dev = dev;
+ pdu->req = req;
+
+ sg_init_table(pdu->usgl.sg, UBI_MAX_SG_COUNT);
+ INIT_WORK(&pdu->work, ubiblock_do_work);
+
+ return 0;
+}
+
+static struct blk_mq_ops ubiblock_mq_ops = {
+ .queue_rq = ubiblock_queue_rq,
+ .init_request = ubiblock_init_request,
+ .map_queue = blk_mq_map_queue,
+};
+
int ubiblock_create(struct ubi_volume_info *vi)
{
struct ubiblock *dev;
@@ -417,13 +397,27 @@ int ubiblock_create(struct ubi_volume_info *vi)
set_capacity(gd, disk_capacity);
dev->gd = gd;

- spin_lock_init(&dev->queue_lock);
- dev->rq = blk_init_queue(ubiblock_request, &dev->queue_lock);
+ dev->tag_set.ops = &ubiblock_mq_ops;
+ dev->tag_set.queue_depth = 64;
+ dev->tag_set.numa_node = NUMA_NO_NODE;
+ dev->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
+ dev->tag_set.cmd_size = sizeof(struct ubiblock_pdu);
+ dev->tag_set.driver_data = dev;
+ dev->tag_set.nr_hw_queues = 1;
+
+ ret = blk_mq_alloc_tag_set(&dev->tag_set);
+ if (ret) {
+ dev_err(disk_to_dev(dev->gd), "blk_mq_alloc_tag_set failed");
+ goto out_put_disk;
+ }
+
+ dev->rq = blk_mq_init_queue(&dev->tag_set);
if (!dev->rq) {
- dev_err(disk_to_dev(gd), "blk_init_queue failed");
+ dev_err(disk_to_dev(gd), "blk_mq_init_queue failed");
ret = -ENODEV;
- goto out_put_disk;
+ goto out_free_tags;
}
+ blk_queue_max_segments(dev->rq, UBI_MAX_SG_COUNT);

dev->rq->queuedata = dev;
dev->gd->queue = dev->rq;
@@ -437,7 +431,6 @@ int ubiblock_create(struct ubi_volume_info *vi)
ret = -ENOMEM;
goto out_free_queue;
}
- INIT_WORK(&dev->work, ubiblock_do_work);

mutex_lock(&devices_mutex);
list_add_tail(&dev->list, &ubiblock_devices);
@@ -451,6 +444,8 @@ int ubiblock_create(struct ubi_volume_info *vi)

out_free_queue:
blk_cleanup_queue(dev->rq);
+out_free_tags:
+ blk_mq_free_tag_set(&dev->tag_set);
out_put_disk:
put_disk(dev->gd);
out_free_dev:
@@ -461,8 +456,13 @@ out_free_dev:

static void ubiblock_cleanup(struct ubiblock *dev)
{
+ /* Stop new requests to arrive */
del_gendisk(dev->gd);
+ /* Flush pending work */
+ destroy_workqueue(dev->wq);
+ /* Finally destroy the blk queue */
blk_cleanup_queue(dev->rq);
+ blk_mq_free_tag_set(&dev->tag_set);
dev_info(disk_to_dev(dev->gd), "released");
put_disk(dev->gd);
}
@@ -490,9 +490,6 @@ int ubiblock_remove(struct ubi_volume_info *vi)
list_del(&dev->list);
mutex_unlock(&devices_mutex);

- /* Flush pending work and stop this workqueue */
- destroy_workqueue(dev->wq);
-
ubiblock_cleanup(dev);
mutex_unlock(&dev->dev_mutex);
kfree(dev);
@@ -620,8 +617,6 @@ static void ubiblock_remove_all(void)
struct ubiblock *dev;

list_for_each_entry_safe(dev, next, &ubiblock_devices, list) {
- /* Flush pending work and stop workqueue */
- destroy_workqueue(dev->wq);
/* The module is being forcefully removed */
WARN_ON(dev->desc);
/* Remove from device list */
--
1.8.4.5

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/